Tagsplanations: Explaining Recommendations Using Tags Jesse Vig Shilad Sen John Riedl Grouplens Research Grouplens Research Grouplens Research University of Minnesota University of Minnesota University of Minnesota [email protected] [email protected] [email protected] ABSTRACT While recommender systems tell users what items they might like, explanations of recommendations reveal why they might like them. Explanations provide many benefits, from im- proving user satisfaction to helping users make better deci- sions. This paper introduces tagsplanations, which are ex- planations based on community tags. Tagsplanations have two key components: tag relevance, the degree to which a tag describes an item, and tag preference, the user’s senti- ment toward a tag. We develop novel algorithms for esti- mating tag relevance and tag preference, and we conduct a Figure 1: Intermediary entities (center) relate user to recommended item. user study exploring the roles of tag relevance and tag prefer- ence in promoting effective tagsplanations. We also examine which types of tags are most useful for tagsplanations. Research shows that explanations help users make more ac- curate decisions [1], improve user acceptance of recommen- ACM Classification Keywords dations [6], and increase trust in the recommender system H.5.3 Information Interfaces and Presentation: Group and [13]. Moreover, studies indicate that users want explana- Organization Interfaces—Collaborative computing; H.5.2 In- tions of their recommendations – a survey of users of one formation Interfaces and Presentation: User Interfaces movie recommender site showed that 86% of those surveyed wanted an explanation feature added to the site [6]. General Terms While many different types of explanation facilities exist, Design, Experimentation, Human Factors they all seek to show how a recommended item relates to a user’s preferences. As Figure 1 illustrates, a common tech- Author Keywords nique for establishing the relationship between user and rec- Explanations, tagging, recommender systems ommended item is to use an intermediary entity. An inter- mediary entity is needed because the direct relationship be- INTRODUCTION tween user and item is unknown, assuming that the user has While much of the research in recommender systems has not yet tried the item. In the Netflix example above, the inter- focused on improving the accuracy of recommendations, re- mediary entity is the list of previously rated movies shown in cent work suggests a broader set of goals including trust, the explanation. The relationship between the user and these user satisfaction, and transparency [16, 1, 13, 6]. A key to movies is that he or she has rated them positively. The rela- achieving this broader set of goals is to explain recommen- tionship between these movies and the recommended movie dations to users. While recommendations tell users what is that other users who liked these movies also liked the rec- items they might like, explanations reveal why they might ommended movie. like them. An example is the “Why this was recommended” feature on Netflix1. Netflix explains movie recommenda- Explanations of recommendations fall into one of three cat- tions by showing users similar movies they have rated highly egories: item-based, user-based, and feature-based, depend- in the past. ing on the type of intermediary entity used to relate the user 1 to the recommended item. In item-based explanations like http://www.netflix.com the Netflix example, a set of items serves as the intermedi- ary entity. User-based explanations utilize other users as in- termediary entities. For example, Herlocker et al. designed Permission to make digital or hard copies of all or part of this work for a explanation that shows a user how other users with simi- personal or classroom use is granted without fee provided that copies are lar taste rated the recommended item [6]. Feature-based ap- not made or distributed for profit or commercial advantage and that copies proaches use features or characteristics of the recommended bear this notice and the full citation on the first page. To copy otherwise, or item as intermediary entities. For example, one movie rec- republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. ommender prototype uses movie features including genre, IUI’09, February 8 - 11, 2009, Sanibel Island, Florida, USA. director, and cast to justify recommendations [14]. Copyright 2009 ACM 978-1-60558-331-0/09/02...$5.00. We present a new type of explanation that uses tags as fea- recommended. Tagsplanations promote understanding by tures, which we call a tagsplanation. The intermediary en- demonstrating how a tag relates to the item and how the tity for tagsplanations is a tag or a set of tags. For example: user relates to the tag. We investigate the role of these two components in helping users understand the recommenda- “We recommend the movie Fargo because it is tagged with tion overall. quirky and you have enjoyed other movies tagged with quirky.” RQ-Effectiveness: What is the role of tag preference and tag relevance in helping users determine if they will like Tags have become increasingly popular on websites such as the item? Delicious2, Flickr3, and Amazon4, and they have many qual- ities that make them useful for explanations. As described in We investigate whether users prefer information about the [5], tags may describe what an item is, what it is about, or relationship between themselves and the tag (tag preference) what its characteristics are – all of which may be useful for explaining a recommendation. Another advantage is that no or between the tag and the item (tag relevance) to make good experts are needed to create and maintain tags, since tags decisions. are applied by the users themselves. Furthermore, tags pro- vide both factual and subjective descriptions [11]. However, RQ-Mood: What is the role of tag preference and tag rel- tags present unique challenges, including issues of tag qual- evance in helping users decide if an item fits their current ity [10] and tag redundancy [5]. mood? We study two aspects of tag-based explanations: the rela- Recommender systems typically do not consider the user’s tionship of the tag to the recommended item, which we call mood when generating recommendations. Explanations pro- tag relevance, and the relationship of the user to the tag, vide users with additional information that can help them which we call tag preference. Tag relevance represents the decide if an item fits their current mood. We investigate the degree to which a tag describes a given item. For example, relative importance of tag preference and tag relevance for consider a tagsplanation for the movie Pulp Fiction that uses revealing mood compatibility. the tag “dark comedy”. In this example, tag relevance would measure how well “dark comedy” describes Pulp Fiction. RQ-Tag-Type: What types of tags are most useful in tags- Tag preference, on the other hand, measures the user’s sen- planations? timent to the given tag, for example how much the user likes or dislikes dark comedies. Tag relevance and tag preference A wide variety of tags may be applied to an item, some of are orthogonal to one another: the former is item-specific which may be more suitable for explanations than others. and the latter is user-specific. As discussed in [10], factual tags identify facts about an item subjective Our design of tagsplanations is motivated by three goals: such as people, places, or concepts, while tags ex- justification, effectiveness, and mood compatibility. Justi- press users’ opinions of an item. We investigate the relative fication is the ability of the system to help the user under- usefulness of factual tags versus subjective tags, and analyze stand why an item was recommended [6]. Justification dif- which specific tags perform best. fers from transparency [16] because justifications may not To answer these research questions, we designed a tagspla- reveal the actual mechanisms of the recommender algorithm. nation feature for the MovieLens movie recommender site5 Tintarev et al. define effectiveness as the ability of the expla- and conducted an online user study. Participants in the study nation to help users make good decisions. Mood compati- viewed 4 types of tagsplanations, each of which handles tag bility is the ability of the explanation to convey whether or preference and tag relevance in a different way. Participants not an item matches a user’s mood. A recent study showed evaluated each tagsplanation based on how well it achieved that users are interested in explanations with mood-related the goals of justification, effectiveness, and mood compati- features [15]. bility. We then analyze the results to determine the roles of In this paper, we investigate the roles of tag relevance and tag relevance and tag preference in promoting these 3 goals. tag preference in tagsplanations. Specifically, we consider Subjects also evaluate specific tags, and we compare the re- the following research questions: sults for subjective tags versus factual tags. RQ-Justification: What is the role of tag preference and RELATED WORK tag relevance in helping users understand their recom- Item-based explanations. Item-based approaches use items mendation? as intermediary entities to relate the user to the recommended item. An example is the “Why this was recommended” fea- Explanations help users understand why a given item was ture on Netflix, which shows users their past ratings for a set of related movies. Similarly, Amazon shows users their past purchases that motivated a recommendation of a new item. 2http://del.icio.us Studies show that item-based explanations improve users’ 3http://www.flickr.com 4http://www.amazon.com 5http://www.movielens.org acceptance of recommendations [6] and help users make ac- many of which may be addressed by using tags. One lim- curate decisions [1].